Improved Factorized Residual U-Net for Biomedical Polyp Segmentation

Colorectal cancer (CRC) is a cancer in the colon or rectum which often begins with the appearance of a polyp. Polyp colorectal is often screened using colonoscopy. While effective, this process is time-consuming. Traditional segmentation approaches for polyp detection face challenges due to the diverse shapes, image quality, and angles in colonoscopy images. Recent advancements in computer vision and deep learning offer promising solutions to improve the efficiency and accuracy of polyp detection. To address these, the authors propose the Improved Factorized Residual U-Net. This model modified U-Net architecture by integrating an improved ResNet backbone and non-bottleneck-1D modules derived from ERFNet, combining the strengths of effective feature extraction and computational efficiency.. Tested on the CVC-ClinicDB dataset, this method achieves an F1-Score of 94.75% and IoU of 90.24%, outperforming existing methods like U-Net++, DoubleUNet, and SSFormer-L in segmentation accuracy. This approach addresses the limitations of existing methods and demonstrates a robust solution for accurate polyp segmentation, potentially revolutionizing CRC screening through enhanced deep learning techniques.